The ground segmentation of 3D LIDAR point cloud with the optimized region merging

This paper represents a additional approach to enhance the result of ground segmentation method with the gathered point cloud from 3D LIDAR. In this segmentation process, the over-segmentation is usually occurred due to the characteristics of 3D LIDAR such as noise, occlusion, and straightness in complex urban environment. In addition, it has a fatal influence on the entire performance of the perception. In this paper, the region merging algorithm for 3D LIDAR point cloud is proposed to integrate overly partitioned ground regions, which are obtained through the region growing algorithm. First, the initial ground is determined by the current vehicle pose, and then the partitioned regions are ordered according to the distance to the vehicle. In this order, both the ground, where the vehicle is able to reach and the respective region are resampled to pairs of the closest edge pixels. If the resampled edge pixels are satisfied with the region merging criterion, the ground region can merge with the compared region and can expand. This process is iterated until all of the partitioned regions are inspected. The proposed region merging algorithm is demonstrated with the labeled simulation data and the real 3D LIDAR data, as compared to the segmentation method without the propsed region merging.